Michele,
you would do this for example if you want to have lesser smoothing
for individual subject's analyses but greater smoothing for group
analyses. I wondered about this issue too recently, and got a very
helpful comment from Matthew Brett on this.
Suppose you have smoothed your data with 8mm kernel,
you can use the following formula to dertermine how to
smooth conimages to result in a smoothing of 12 mm. However,
there are some caveats that are described below. You may
be better off re-doing your glm with differently smoothed
raw data.
12 = sqrt(smoothing_needed.^2 + 8.^2)
->
smoothing_needed = sqrt(122-8.^2) = 8.94mm
_but_ you need to be very careful, because the contrast images have NaNs
in the areas outside the brain. If you smooth the images like this,
everything within the smoothing kernel that hits an NaN, becomes a NaN.
This means that smoothing will shave a smoothing-kernel thickness off
your brain. So, you would probably want to replace the NaNs in the
contrasts images with zeros. This still doesn't give you quite what you
want, because now you have a sharp brain -> 0 edge which could cause
problems, and you have lost the within-brain definition that the NaNs
gave you. For this reason, if you can bear it, I would recommend
redoing the GLMS at the first level instead...
Michele Wessa wrote:
>I have a question concerning the smoothing of contrast images. Obviously
>there are some groups who smooth the contrast images after already having
>smoothed the normalised images (for exemple smooth the normalised images
>with a Kernel of 6mm and re-smooth the contrast images also with a kernel
>of 6mm). I am not sure what are the reasons for doing this? I asked myself
>if it is helpful to use this procedure, if I don't want to use a wide
>smoothing for the normalised data because I want to see for example small
>structures (as the amygdala) and then to re_smooth the original images to
>reveal better random effects statistics (I have no idea if this makes
>sense)?
>
>Happy to have suggestions and explanations on this topic -
>
>Michèle
>
>
>
--
Christian Fiebach, PhD
D'Esposito Neuroimaging Laboratory
Department of Psychology and
Helen Wills Neuroscience Institute
UC Berkeley
132 Barker Hall
Berkeley, CA 94720-3190
Tel.: 1-510-642-2839
mail: mailto:[log in to unmask]
web: http://fiebach.org
|